Embedded evaluators can check at the level of nuts and bolts whether an AI company is actually following the training, deployment, operational, and safeguards practices they claim to be following.
Virtue lens: When AI is becoming as powerful as it is, at the speed that it is, following safeguards are the only ways we can make sure it is somewhat safe. It isn't even just about being responsible or honest. If a program gaining this kind of traction has no sort of training, safeguards, guidelines, etc., how will it continue to function without falling apart or harming someone. So having some kind of base level of safety and operational practices should, honestly, already be something that was thought of and put in place.